How to visualize input Dataloader after Augmentation?

[chokyungjin]

Hi! I have a question your Dataloader.
Your yolov5 is soo difficult to me to analyze..
How to visualize input Dataloader after Augmentation?
Before I training, I wonder how images come out in the dataloader.

Thanks!

[Ownmarc]

Trust me, Glenn’s code is really easy to read and understand compared to many repo in ML.

You can check the images created in the folder ‘’runs/exp0/‘’ they will show you 3 first batches when starting a training. Else you could simply create a dataloader object in a notebook or something using the repo’s code and load image one after an other and view them however you want by writing to disk or showing them with pyplot

[DacDinh147]

You have to look in the LoadImagesAndLabels(Dataset) class. This class return a tuple (torch.from_numpy(img), labels_out, self.img_files[index], shapes).

Here torch.from_numpy(img) is a tensor with shape [batch size, channels, w, h] and labels_out is a list of item. Each item of labels_out represents a bounding box annotation for each image and is a list of bounding box (x_center, y_center, w, h) corresponding to the object in the specific image. Because the bounding box coordinates (x_center, y_center, w, h) was normalized, so you have to multiply it with the width and height of the image to visualize it xc, yc, w, h = xc*w_img, yc*h_img, w*w_img, h*h_img

Visualization with batch size=6

[glenn-jocher]

We display this clearly in the notebook, suggest you start from there:
https://colab.research.google.com/github/ultralytics/yolov5/blob/master/tutorial.ipynb#scrollTo=DLI1JmHU7B0l

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[buimanhlinh96]

Hi @glenn-jocher, how can i add albumentations in torch library for augmentations in DataLoader?

[glenn-jocher]

@buimanhlinh96 you can customize the trainloader here:

yolov5/utils/datasets.py

Line 328 in 7220cee

class LoadImagesAndLabels(Dataset): # for training/testing

[Auth0rM0rgan]

@glenn-jocher, Could you please provide more details that how to use albumentations in the DataLoader??

[glenn-jocher]

@Auth0rM0rgan we don't have an integration with albumentations. YOLOv5 augmentation hyperparameters are set here:

yolov5/data/hyp.scratch.yaml

Lines 22 to 33 in f6b3c96

	hsv_h: 0.015 # image HSV-Hue augmentation (fraction)
	hsv_s: 0.7 # image HSV-Saturation augmentation (fraction)
	hsv_v: 0.4 # image HSV-Value augmentation (fraction)
	degrees: 0.0 # image rotation (+/- deg)
	translate: 0.1 # image translation (+/- fraction)
	scale: 0.5 # image scale (+/- gain)
	shear: 0.0 # image shear (+/- deg)
	perspective: 0.0 # image perspective (+/- fraction), range 0-0.001
	flipud: 0.0 # image flip up-down (probability)
	fliplr: 0.5 # image flip left-right (probability)
	mosaic: 1.0 # image mosaic (probability)
	mixup: 0.0 # image mixup (probability)

You are free to modify the dataloader also as you see fit here:

yolov5/utils/datasets.py

Line 328 in 7220cee

class LoadImagesAndLabels(Dataset): # for training/testing

[Auth0rM0rgan]

Hey @glenn-jocher, Thanks for the quick reply!

I know about the YOLOv5 augmentation but I'm trying to add some blur augmentations that are not in the yolov5 augmentation such as MotionBlur, Gaussian Blur, and ... but I have difficulty to add these augmentations inside YOLOv5. I will appreciate if you can help me even a little...

Gracias!

[glenn-jocher]

@Auth0rM0rgan ah I see. Well you can access the raw image as a numpy array img here after it's already passed through the YOLOv5 augmentations, so I would just insert any augmentations here. Be careful to correctly modify labels correspondingly as well if your augmentations require it though.

yolov5/utils/datasets.py

Lines 526 to 528 in f6b3c96

	if mosaic:
	# Load mosaic
	img, labels = load_mosaic(self, index)

[Auth0rM0rgan]

@glenn-jocher Thanks! made it work. Now I can apply a set of different kinds of augmentations at pixel-level and spatial-level!

[glenn-jocher]

@Auth0rM0rgan hey cool! Maybe you could consider submitting a PR for general Albumentations integration? I know its a popular tool so others might find that useful.

[Auth0rM0rgan]

@glenn-jocher, Sure! will do it during the weekend!

[glenn-jocher]

@Auth0rM0rgan @buimanhlinh96 see PR #3882 for a proposed automatic Albumentations integration.